Fun technology tutorials for a cyberpunk future. We share what we learn about AI content creation, open-source tools, and creative coding — because knowledge wants to be free.
Every video on our Instagram is made entirely with AI. No camera, no studio, no film crew. Just text prompts, generative video models, and a Python script that syncs everything to music. Here's how we do it.
Our primary video generation model. Kling excels at cinematic camera movements, consistent character appearance across clips, and photorealistic output. We use it for hero shots and any scene that needs to look indistinguishable from real footage.
Our go-to for stylized and artistic content. Runway handles abstract visuals, transitions, and creative effects better than any other model. We use it when we want a video to feel more like digital art than photography.
Google's latest video model brings exceptional prompt adherence and smooth motion. We use Veo for scenes requiring precise actions (a cat turning its head, walking toward camera) where following the prompt exactly matters most.
Every video starts with a concept. We write a shot list describing each scene: the subject, camera angle, lighting, mood, and movement. This becomes the prompt sheet. A 60-second video typically needs 8-12 individual clips, each 4-8 seconds long.
We feed each prompt into the appropriate AI model. Kling for photorealistic hero shots, Runway for stylized sequences, Veo for precise movements. Each prompt usually generates 3-5 variations. We cherry-pick the best take for each shot, looking for visual consistency, motion quality, and adherence to the storyboard.
Pro tip: generating multiple variations is key. AI video models are non-deterministic — the same prompt produces different results each time. We typically generate 3-5 takes per shot and pick the best one. Think of it like film production: you wouldn't use the first take of every scene.
This is where it gets technical. We wrote a Python script that analyzes a music track, detects beats using librosa, and automatically cuts between clips on beat drops. The script uses moviepy for video editing and handles crossfades, speed ramps, and color grading. The result: cinematic, music-synced videos that feel professionally edited.
Final touches include color grading for visual consistency across clips, adding text overlays, adjusting audio levels, and exporting at the right resolution for each platform. Instagram Reels, TikTok, and YouTube Shorts all have different optimal specs. We batch-export from the Python pipeline to hit all platforms in one run.
Why Python over Premiere/DaVinci?
Repeatability. Once the script is written, creating a new video takes minutes instead of hours. Change the music track, swap in new clips, and the script handles all the timing automatically. It's the cyberpunk way — automate everything.
"We're not filmmakers. We're not video editors. We're cat breeders who learned Python and figured out how to make AI do the heavy lifting. If we can do it, you can too."
— Persian PunksTurn any photo into anime-style art using PyTorch and AnimeGAN — a lightweight generative adversarial network that transforms real images into anime versions. No LLM required, no API calls, no subscription fees. Just open-source machine learning running on your own hardware.
AnimeGAN is a generative adversarial network trained specifically on anime art styles. It learns the visual patterns of anime (bold outlines, flat colors, stylized shading) and applies them to real photographs. The model runs entirely locally — your images never leave your machine.
Python 3.8+, PyTorch, and a GPU with at least 4GB VRAM (CPU works too, just slower). The AnimeGAN model weights are freely available and only around 8MB. No cloud API, no tokens, no ongoing costs. Install the dependencies, download the weights, and you're running in under five minutes.
We built a browser-based demo so you can test AnimeGAN without installing anything. Upload a photo, pick an anime style, and see the result in seconds. It runs client-side in your browser — your images stay private.
openpilot is an open-source driver assistance system developed by comma.ai. It provides adaptive cruise control and lane centering for over 300 supported car models — essentially giving your existing car semi-autonomous driving capabilities for a fraction of the cost of factory ADAS upgrades.
openpilot provides two core features: adaptive cruise control (maintains speed and following distance) and lane centering (keeps your car centered in the lane). On supported cars, it handles highway driving with minimal driver input. It uses a camera-first approach, similar to Tesla's vision system, but fully open source.
comma.ai sells the comma 3X, a dedicated device that mounts behind your rearview mirror and plugs into your car's OBD-II port. It runs openpilot, records driving data, and processes everything on-device. The entire system is designed to be installed in under 30 minutes with no permanent modifications to your car.
Open-source self-driving software is the most cyberpunk technology on the planet right now. A community of hackers and engineers building driver assistance that rivals billion-dollar OEM systems — and giving it away for free. We use openpilot daily on our drives across Texas, and it is genuinely impressive.
"The comma 3X is the best driving companion we own, and openpilot gets better with every update. It's open source, it's community-driven, and it turns a regular car into something that feels like the future."
— Persian PunksA Python toolkit that scrapes your Counter-Strike 2 Steam inventory, pulls live market prices, fetches lifetime price history, and generates an interactive dashboard so you can see exactly what your skins are worth — and how that's changed over time. Open source, no subscriptions, runs locally.
Fetches your full CS2 inventory directly from Steam and pulls current Steam Market prices for every item. Data is cached locally as JSON so repeat runs are fast and don't hammer the API.
Retrieves lifetime price history from the Steam Market for each item and renders interactive line charts. See how a skin's value has trended over months — useful for spotting buy and sell windows.
Generates a self-contained HTML dashboard with tabbed views, sortable columns, and filtering. No backend required — open the file in any browser. Also exports Markdown reports for quick reference.
"CS2 skins are genuinely volatile assets. We built this because we wanted a clear picture of what our inventory was actually worth — and whether holding or selling made sense. Now we know."
— Persian PunksOn-chain data is the blockchain's heartbeat — raw, unfiltered signal from the network itself. Unlike price charts, on-chain metrics reveal what's actually happening: who's moving coins, how much supply is dormant, whether the market is overheated, how deeply ETH is embedded in DeFi. We built a live dashboard so you can see it all in one place.
Price is what someone is willing to pay right now. On-chain data is what's actually happening. MVRV tells you if the market is overheated. Active address trends signal network growth or decline. Long-term holder supply shows conviction. It's the difference between the headline and the story.
Bitcoin's on-chain story is security and scarcity. Hash rate at all-time highs means the network has never been more secure. Long-term holder supply above 74% means most BTC hasn't moved in over a year — classic accumulation behavior. MVRV below 2 historically signals undervaluation.
Post-merge Ethereum burns ETH with every transaction via EIP-1559, making high-activity periods deflationary. ~28% of supply is now staked and earning yield. The DeFi TVL locked on-chain ($45B+) represents real economic activity that simply doesn't exist in Bitcoin's design space.
Market Value to Realized Value. Realized value is the price each coin last moved on-chain, summed across the entire supply. MVRV below 1 means most holders are underwater — historically a bottom signal. Above 3.5 is historically overheated. Between 1–2 is the fair-to-undervalued range.
Cardano's on-chain story is participation. With ~65% of supply staked across ~3,000 pools, it has the highest staking ratio of the three and one of the most decentralized validator sets in proof-of-stake. The ecosystem is smaller but the architecture reflects a deliberate philosophy: broad participation over raw throughput.
Live market data via CoinGecko API (price, volume, market cap, 7-day history). On-chain metrics sourced from Glassnode, IntoTheBlock, and Cardanoscan — approximate values updated periodically. Not financial advice.
"We don't trade. We're cat breeders. But we do hold crypto, we accept it for merch now, and we built this because we wanted to understand what we're actually holding. On-chain analysis is how you look past the noise."
— Persian PunksMore tutorials and deep-dives in the pipeline. Follow us for updates.
Run large language models locally on your own hardware. No cloud, no API keys, no data leaving your machine. We'll cover setup, model selection, and practical use cases.
A comprehensive guide to creating consistent characters and scenes with Flux, Midjourney, and Stable Diffusion. Prompt engineering, LoRA training, and workflow automation.
The tools and services we use to stay private online. From ProtonMail to VPNs, encrypted messaging, and privacy-respecting alternatives to big tech services.
We share tutorials, AI experiments, behind-the-scenes content creation, and of course — Persian cats. The cyberpunk cat community is growing.